evo 🤖 Agent Open source

What it does

evo turns a codebase into an automated optimisation loop. You point it at a repository, it works out what can be measured, sets up a benchmark, and then runs rounds of experiments in which coding agents try changes, keep the ones that improve the score and discard the rest. It builds on the autoresearch idea of an agent hill-climbing its own results, but replaces the single greedy path with a tree search and parallel workers.

What is inside

Two commands drive it: discover explores the repo, asks what to optimise and instruments the evaluation, and optimize runs the loop. Each subagent works in its own git worktree, reads shared failure traces and discarded hypotheses, and can iterate on its branch. Between rounds a frontier strategy (argmax, top-k, epsilon-greedy, softmax or Pareto per task) picks which branch to extend. Gates are pass or fail commands such as a test suite that every experiment must clear, which stops the search from gaming the metric. A local dashboard shows experiments, and runs can execute locally, over SSH or on cloud sandboxes such as Modal, E2B, Daytona, AWS or Azure.

Works with

Claude Code, Codex, Cursor and OpenCode, plus other hosts including Kimi, Hermes, Pi and OpenClaw. The invocation syntax differs per host.

How to install

Install the CLI, then add the plugin and hooks for your host:

uv tool install evo-hq-cli
evo install <host>

Maintenance and safety

Apache-2.0 licensed with tests in CI and versioned releases; the README documents upgrade and migration steps in detail. evo sends anonymous telemetry by default, which can be switched off with evo telemetry off. By default it runs unattended with parallel subagents editing code, so set up gates before letting it loose, and keep an eye on compute and model costs when using cloud backends.

Who should use it

Engineers with a measurable target, such as parser speed, benchmark accuracy or latency, who want coding agents to explore many candidate changes systematically. It is less useful for work without a clear metric.